The aim of the paper is to build a generalised outbreak detection methodology for Poisson counts data. Efficient multivariate spatio-temporal outbreaks detection algorithms are not currently available in the literature. This paper offers a recursive partitioning approach for identifying unusually higher counts than expected in a clustered multivariate space. The approach is applied to the problem of early detection of unusually high vehicle crashes. Multivariate clustered outbreaks are searched for in the dimensions of age and gender of the person causing the crash, the vehicle type, the road type, the road movement during the crash, and the geographical location of the crash. The focus is on persistent outbreaks which are more likely to benefit from feedback control actions.